TS: port CPU and NMC versions to dfatool

4 files changed, 42 insertions, 16 deletions

diff --git a/TS/baselines/cpu/Makefile b/TS/baselines/cpu/Makefile
index 69121ed..b0b9a86 100644
--- a/TS/baselines/cpu/Makefile
+++ b/TS/baselines/cpu/Makefile
@@ -1,4 +1,3 @@
-
 all: streamp_openmp
 
 streamp_openmp: streamp_openmp.cpp tools.cpp
@@ -13,10 +12,15 @@ streamp_openmp_O2: streamp_openmp.cpp tools.cpp
 run: streamp_openmp
 	./streamp_openmp inputs/randomlist33M.txt 256
 
+# may need OMP_NUM_THREADS=32 (≈ tinos) -- does not work with 88 threads @ ios
+
 run_O0: streamp_openmp_O0
 	./streamp_openmp_O0 inputs/randomlist33M.txt 256
 
 run_O2: streamp_openmp_O2
 	./streamp_openmp_O2 inputs/randomlist33M.txt 256
 
+clean:
+	rm -f streamp_openmp streamp_openmp_O0 streamp_openmp_O2
+
 .PHONY: all run run_O0 run_O2 clean
diff --git a/TS/baselines/cpu/run-opti.sh b/TS/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..9135393
--- /dev/null
+++ b/TS/baselines/cpu/run-opti.sh
@@ -0,0 +1,15 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+for i in $(seq 1 50); do
+	( make run_O0 || OMP_NUM_THREADS=32 make run_O0 ) | sed 's/CPU/CPU O0/'
+done | tee "${HOST}-O0.txt"
+
+for i in $(seq 1 50); do
+	( make run_O2 || OMP_NUM_THREADS=32 make run_O2 ) | sed 's/CPU/CPU O2/'
+done | tee "${HOST}-O2.txt"
diff --git a/TS/baselines/cpu/run.sh b/TS/baselines/cpu/run.sh
index 7d8fe37..254ca15 100755
--- a/TS/baselines/cpu/run.sh
+++ b/TS/baselines/cpu/run.sh
@@ -2,15 +2,22 @@
 
 set -e
 
+HOST="$(hostname)"
+
+echo $HOST
+
+(
+
 echo "prim-benchmarks TS CPU (dfatool edition)"
 echo "Started at $(date)"
 echo "Revision $(git describe --always)"
 
 # input size depends on file -> strong scaling only
 
-make
+make -B
 for i in $(seq 1 10); do
-	for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+	for nr_threads in 88 64 44 1 2 4 6 8 12 16 20 24 32; do
 		OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./streamp_openmp inputs/randomlist33M.txt 256 || true
 	done
 done
+) | tee "${HOST}-explore.txt"
diff --git a/TS/baselines/cpu/streamp_openmp.cpp b/TS/baselines/cpu/streamp_openmp.cpp
index 94f110f..cc970e2 100644
--- a/TS/baselines/cpu/streamp_openmp.cpp
+++ b/TS/baselines/cpu/streamp_openmp.cpp
@@ -290,7 +290,7 @@ int main(int argc, char* argv[])
   std::cout << "///////////////////////// STREAMP //////////////////////////" << std::endl;

   std::cout << "############################################################" << std::endl;

   std::cout << std::endl;

-  std::cout << "[>>] Reading File..." << std::endl;

+  //std::cout << "[>>] Reading File..." << std::endl;

 

   /* Read time series file */

   tstart = std::chrono::high_resolution_clock::now();

@@ -329,7 +329,7 @@ int main(int argc, char* argv[])
   std::cout << std::endl;

 

   // Preprocess, statistics, get the mean and standard deviation of every subsequence in the time series

-  std::cout << "[>>] Preprocessing..." << std::endl;

+  //std::cout << "[>>] Preprocessing..." << std::endl;

   tstart = std::chrono::high_resolution_clock::now();

 

   tprogstart = tstart;

@@ -337,11 +337,11 @@ int main(int argc, char* argv[])
 

   tend = std::chrono::high_resolution_clock::now();

   time_elapsed = tend - tstart;

-  std::cout << "[OK] Preprocess Time:         " << std::setprecision(std::numeric_limits<double>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

-  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_preproc_MBps=%f throughput_preproc_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

+  //std::cout << "[OK] Preprocess Time:         " << std::setprecision(std::numeric_limits<double>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

+  printf("[::] TS CPU | n_threads=%d e_type=%s n_elements=%d | throughput_preproc_MBps=%f throughput_preproc_MOpps=%f", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

 

   //Initialize Matrix Profile and Matrix Profile Index

-  std::cout << "[>>] Initializing Profile..." << std::endl;

+  //std::cout << "[>>] Initializing Profile..." << std::endl;

   tstart = std::chrono::high_resolution_clock::now();

 

   profile          = new DTYPE[ProfileLength];

@@ -354,8 +354,8 @@ int main(int argc, char* argv[])
 

   tend = std::chrono::high_resolution_clock::now();

   time_elapsed = tend - tstart;

-  std::cout << "[OK] Initialize Profile Time: " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

-  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_init_MBps=%f throughput_init_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

+  //std::cout << "[OK] Initialize Profile Time: " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

+  printf(" throughput_init_MBps=%f throughput_init_MOpps=%f", timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

 

   // Random shuffle the diagonals

   idx.clear();

@@ -366,15 +366,15 @@ int main(int argc, char* argv[])
     std::random_shuffle(idx.begin(), idx.end());

 

   /******************** SCRIMP ********************/

-  std::cout << "[>>] Performing STREAMP..." << std::endl;

+  //std::cout << "[>>] Performing STREAMP..." << std::endl;

   tstart = std::chrono::high_resolution_clock::now();

 

   streamp();

 

   tend = std::chrono::high_resolution_clock::now();

   time_elapsed = tend - tstart;

-  std::cout << "[OK] STREAMP Time:            " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

-  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_streamp_MBps=%f throughput_streamp_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

+  //std::cout << "[OK] STREAMP Time:            " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

+  printf(" throughput_streamp_MBps=%f throughput_streamp_MOpps=%f", timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

 

   // Save profile to file

   //std::cout << "[>>] Saving Profile..." << std::endl;

@@ -388,9 +388,9 @@ int main(int argc, char* argv[])
 

   // Calculate total time

   time_elapsed = tend - tprogstart;

-  std::cout << "[OK] Total Time:              " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

-  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_total_MBps=%f throughput_total_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

-  std::cout << std::endl;

+  //std::cout << "[OK] Total Time:              " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;

+  printf(" throughput_MBps=%f throughput_MOpps=%f\n", timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));

+  //std::cout << std::endl;

 

   delete profile;

   delete profileIndex;